Suǒyǐ ‘so’, they are different: an integrated subjectivity account of Mandarin RESULT connectives in conversation, microblog and newspaper discourse

3.1 Corpus, data type and collection

Our data stem from three corpora: newspaper articles, spontaneous conversations, and microblog messages.^[5] These materials were selected because they are reasonably representative of written, spoken, and social media discourse in terms of the production and reception processes and potentially the variation of subjectivity.

The newspaper articles were collected from Rénmín Rìbào ‘People’s Daily’ (time span: 1991–1996). It is part of the Mandarin Chinese News Text corpus LDC95T13 (Wu 1995). Microblog data were taken from Sina Weibo ‘Sina Microblog’.^[6] Akin to the social networking applications of Facebook and Twitter, Sina Weibo is the most popular microblogging platform in China. Our data are taken from the posts in the year 2013, which is a sub-corpus of BCC (Xun et al. 2016).

Incorporated in the spoken corpus are conversations over telephone and in TV interview/talk-show programs. The telephone data consist of two parts. One is from LDC2005S15/T32 (Fung et al. 2005), which contains 897 calls between Mandarin speakers from Mainland China (about 150 h’ audio and transcripts of 2,030,730 Chinese characters). The other part is from LDC96S34/T16 (Canavan and Zipperlen 1996; Wheatley 1996), which contains 120 calls between family members from Mainland China (about 18.3 h audio and transcripts of 303,906 Chinese characters). TV interview/talk-show conversations are extracted from four programs: Lǔyù yǒu yuē ‘A Date with Luyu’, Shíhuà shíshuō ‘Tell It like It Is’, Qiāngqiāng sānrénxíng ‘Behind the Headlines’, and Xīnwén huìkètīng ‘People in the News’. These data stem from two corpus resources. One is LDC, from which we obtained in total 93 h’ audio and corresponding transcripts of 1,199,777 Chinese characters (Glenn et al. 2013, 2014, 2015, 2016; Walker et al. 2013, 2014, 2015, 2016). The other is MLC (2005), from which we made use of the transcripts for the programs Lǔyù yǒu yuē (1,994,244 Chinese characters) and Qiāngqiāng sānrénxíng (5,903,852 Chinese characters).

From each corpus, 50 constructions marked with kějiàn, suǒyǐ, yīncǐ and yúshì, respectively, were randomly collected for further analysis (600 in total). The tokens kějiàn and yúshì do not function only as discourse connectives. Kějiàn can also be a verbal phrase ‘can see’ or an adjective phrase ‘visible’; yúshì very often occurs in phrases such as děngyú shì or xiāngdāngyú shì ‘be equal to’. For this reason, special attention was paid to make sure that only causal connective uses of kějiàn and yúshì were collected. Table 1 presents an overview of the complete datasets and the corpus information.

3.2 Model of analysis

In line with the integrated notion of subjectivity introduced in Section 2, we analyzed each example in terms of three subjectivity features and followed a three-step procedure in annotation. The first is to annotate the domain of the causal relation, which is essentially related to the propositional attitude expressed in the consequent, i.e., the end status of the causal event (referred to as PropAtt hereafter). The second is to identify the subject responsible for the causal interpretation of the relation, i.e., the identity of the SoC (IdSoC). The third step is to annotate the linguistic realization of the SoC (hereafter LingReal-SoC). Each feature contains several categories that differ in their degree of subjectivity (Table 2). Subsequently, the analysis tests whether there are systematic co-occurring patterns between each connective and the subjectivity features, and whether or not the patterns remain robust across corpora.

The basis of this model was built up in previous studies on Dutch causal connectives (Sanders and Spooren 2015; Spooren et al. 2010) and Mandarin causal connectives (Li et al. 2013, 2016; Xiao et al. 2021). The major difference is that we have merged the domain of the relation and the propositional attitude of the result segment due to the correlation between these two features in the model (cf. Table 3).

3.3 Inter-coder agreement and objectivity bias

The codebook for the present study has been tested and proved effective in the study of Mandarin reason connectives (Xiao et al. 2021). Of the 600 instances analyzed in the present study, 180 (30%) were annotated separately by the first two authors, who are native speakers of Mandarin Chinese. The inter-coder agreement (Cohen’s kappa) was found to be substantial: PropAtt κ = 0.80; IdSoC κ = 0.87; LingReal-SoC κ = 0.83. All the discrepancies were discussed afterwards. In the end, complete agreement was reached upon IdSoC and LingReal-SoC; seven cases regarding PropAtt remained in dispute.

To settle the disputes, we referred to the context for justification. Four of the disagreements were related to the interpretation of Q as an epistemic conclusion or as a mental act. The difference between the two categories can be very subtle. By nature, a mental act is distinctive from a conclusion in that the former implies a change of mental state or expresses an activity taking place within the mental world of the SoC, whereas the latter expresses an opinion or the state of mind itself. For the subtleties, see (16), an example from the spoken corpus, for which the preceding context is provided, and Example (17), from the microblog corpus.

The ambiguity is caused by xiǎng ‘want’ in the predicate, which, depending on the context, can indicate both a mental act and an opinion. In (16), “he then wanted to become a police officer” indicates the change of his mind (from “working hard to gain the employees’ recognition” to “considering to take a different job”). This interpretation is also implied by the condition presented in the context: all his efforts failed to change other people’s opinion of him as a son from a wealthy family, and he felt frustrated with it. Moreover, this is also a typical instance of a narrative causal relation expressed by yúshì: an event takes place under some circumstances of another. The conjunction adverb jiù ‘then’ also implies an act of reasoning. Therefore, we annotated it as mental act. However, “I want to go slowly” in (17) indicates the SoC’s opinion. A natural paraphrase is: “we come from […] and head for […]”, so my opinion is “to go slowly”. Hence, we annotated it as judgment.

The remaining three disputes were over an intentional or unintentional reading of the causal event. If the context did not help, or if there was lack of context (especially in microblogs), we resolved the dispute by opting for the unintentional reading. Example (18) is a microblog message in our corpus. Due to lack of context, it remains unclear whether the blogger’s notice of “the change of numbers” means that he/she intentionally pays close attention to the number (in which case Q expresses an act), or that the change of number inevitably catches his/her attention (in which case Q expresses a fact). In such cases, we opted for the objective interpretation: fact.

Note that in the complete dataset of the 600 instances, only 10 cases in total remain ambiguous. We consistently took the objectivity bias (i.e., chose an objective reading) for the coding of these cases.

4.1 Propositional attitude of the consequent

The frequencies and percentages of the five PropAtt categories observed in each corpus and in the relations signaled with each connective are presented in Figures 1 and 2 below, respectively.

Figure 1:

Frequencies and percentages of each PropAtt category observed in the corpus of Newspaper (N), Conversation (C) and Microblog (M).

Figure 2:

Frequencies and percentages of each PropAtt category co-occuring with kějiàn, suǒyǐ, yīncǐ and yúshì.

The log-linear analysis resulted in a model containing two significant two-way interactions. The fit of the model was substantial, χ² (30, N = 600) = 13.65, p = 0.995. The first interaction was between Corpus and PropAtt, χ² (8, N = 600) = 24.07, p = 0.002. The second was between Connective and PropAtt, χ² (12, N = 600) = 451.95, p < 0.001. The interactions were followed up by cross-table analyses summarized in Tables 4 and 5.

The adjusted residuals in Table 4 show that the significant interaction between Corpus and PropAtt was exclusively due to the corpus-sensitive distribution of Speech act. A significantly high frequency was observed in Microblog (a.r. = 4.4). Example (19) illustrates a typical speech act instance from the microblog data.

The adjusted residuals in Table 5 show that the co-occurrence pattern of Connective and PropAtt varies a lot from one connective to another. Kějiàn expresses predominantly Judgments (149 out of 150, a.r. = 15.8). Yúshì behaves just the opposite. It barely expresses any subjective category of Speech act (a.r. = −3.2) or Judgment (a.r. = −12.2), but mainly the objective ones: Physical act (a.r. = 11.3) and Facts (a.r. = 3.1). Suǒyǐ and yīncǐ appear less extreme. Generally, suǒyǐ is more on the subjective side, as witnessed by the high adjusted residuals for Speech act (a.r. = 4.5).

In Example (19) above, “be sure to be patient […]” illustrates the typical speech act relation introduced by suǒyǐ. The Q-segment “there are often mice in this house” in Example (20) is a claim (Judgment) by the SoC, which is almost the only type of PropAtt found in relations signaled by kějiàn. Yīncǐ tends to express factual Q-segments. Physical act occurs most frequently in yúshì relations, as exemplified by (10), repeated as (21): “came to find Běiyīng Wēn […]”. Example (22) illustrates a factual Q-segment in yúshì cases: the appearance of “profiteers” is presented as a natural consequence due to the “rise in prices” on the tulip market.

4.2 Identity of the SoC

The frequencies and percentages of the three types of SoC identity observed in each corpus and in the relations marked with each connective are presented respectively in Figures 3 and 4 below.

Figure 3:

Frequencies and percentages of each IdSoC category observed in Newspaper (N), Conversation (C) and Microblog (M).

Figure 4:

Frequencies and percentages of each IdSoC category co-occurring with kějiàn, suǒyǐ, yīncǐ and yúshì.

The log-linear analysis based on the identity of the SoC resulted again in a model containing two significant two-way interactions. One was between Corpus and IdSoC, χ² (4, N = 600) = 30.98, p < 0.001. The other was between Connective and IdSoC, χ² (6, N = 600) = 197.30, p < 0.001. The fit of the model was adequate, χ² (18, N = 600) = 14.32, p = 0.708. The following crosstab analyses provide us an insight into the two interactions.

The adjusted residuals in Table 6 show that the interaction between Corpus and IdSoC was largely due to the corpus-sensitive distribution of Speaker SoC and Character SoC. The former was relatively rare in Newspaper (a.r. = −3.4). The latter occurred relatively often in Newspaper (a.r. = 5.4), but rarely in Microblog (a.r. = −3.9). Example (21) from our written data exemplifies a character SoC, i.e. “the villagers”, undertaking the physical act “came to find Běiyīng Wēn […]”.

The adjusted residuals in Table 7 cast light on the interaction between Connective and IdSoC. All intances with kějiàn have a speaker SoC (n = 150, a.r. = 10.8). By contrast, yúshì co-occurred relatively rarely with Speaker SoC (a.r. = −8.4) and mainly with Character SoC (a.r. = 7.2) and No SoC (a.r. = 3.1). As witnessed by the adjusted residuals, suǒyǐ relations appeared neutral, although the absolute frequencies of Speaker SoC were rather high (98 out 150). In yīncǐ relations, the most subjective category Speaker SoC occurred less than expected (a.r. = −2.9), and the objective category No SoC almost reached significance (a.r. = 2.8 vs. 2.87).

Example (20) in Section 4.1 illustrates a typical kějiàn relation conceptualized by a speaker SoC: the speaker draws the conclusion that “there are often mice in this house”. Typical yúshì relations are exemplified via (21) with a character SoC, “the villagers” who “came to find Běiyīng Wēn […]”, and (22) with a factual Q-segment describing a real-world event, i.e., “a surge of profiteers on the market of tulips”. A yīncǐ instance with no SoC is Example (11), repeated below as (23), which introduces the fact of an old joke exiting in the local area.

4.3 Linguistic realization of SoC

Presented in Figure 5 are the frequencies and percentages of the four LingReal-SoC categories co-occurring with every connective in, respectively, the corpus of newspaper, conversation and microblog.

Figure 5:

Frequencies and percentages of each LingReal-SoC category co-occuring with kějiàn, suǒyǐ, yīncǐ and yúshì in Newspaper (N), Conversation (C) and Microblog (M).

The log-linear analysis generated a model containing a significant three-way interaction: Connective*LingReal-SoC*Corpus, χ² (18, N = 600) = 36.51, p = 0.006, indicating that the relation between Connective*LingReal-SoC was moderated by Corpus. Table 8 below reveals the behaviors of the connectives, in Newspaper (p < 0.001, Fisher’s exact test), Conversation (p < 0.001, Fisher’s exact test), and Microblog (χ² (9, N = 200) = 70.82, p < 0.001), respectively.^[8]

Across all the three corpora, kějiàn co-occurred predominantly with Implicit, rarely with Pro-drop and Absent; the pattern with Explicit varied from being rare in Newspaper (count: 2; a.r. = −5.6) and Conversation (count: 6; a.r. = −5.0) to not so rare in Microblog (count: 11, a.r. = −0.3). In contrast to kějiàn, in all three corpora, yúshì rarely occurred with Implicit. The pattern regarding Explicit and Pro-drop varied across corpora: relatively many Explicit in Newspaper (count: 31; a.r. = 4.2), and relatively many Pro-drop in both Newspaper and Microblog (respectively, a.r. = 4.9 and 3.8). Suǒyǐ and yīncǐ appeared relatively neutral in all corpora.

Exemplified in (24) is kějiàn co-occurring with an explicit SoC in microblog; Examples (25) and (26) demonstrate the use of yúshì with an explicit SoC in newspaper and pro-drop in microblog.

4.4 Relative importance of the three subjectivity features

To measure the strength of each of the three subjectivity features in distinguishing between the connectives under study, we performed the conditional inference trees and random forest analyses using the packages party and Rling in R (Levshina 2015: 291–300; R Core Team 2017). These are two non-parametric tree-structure models of regression and classification that are appropriate for our dataset, as it is relatively small in size and has multiple categorical predictors. The conditional inference trees analysis resulted in several splits (graphically represented as a tree) reflecting important predictors in the model (only significant splits are presented). The analysis also provides p values representing the confidence level of each split.

The conditional inference tree below (Figure 6) presents the decision rules for distinguishing kějiàn, suǒyǐ, yīncǐ, and yúshì.^[9] The first split involves PropAtt, which set apart Judgment from the rest. A further split within Judgment brought LingReal-SoC into play (Node 7). This node includes 52 cases of Explicit Judgment (Node 8) which contain mainly suǒyǐ (40%) and kějiàn (35%) and a moderate amount of yīncǐ (20%). The 212 cases of Implicit/Pro-drop Judgment (Node 9) contain predominantly kějiàn (60%); the rest are suǒyǐ and yīncǐ (each circa 20%).

Figure 6:

Conditional Inference Tree predicting the occurrence of four RESULT connectives on the basis of PropAtt, LingReal-SoC, IdSoC.

The second split (Node 2) again related to PropAtt and set apart Speech act from Mental/Physical act and Fact, resulting in a bin containing mainly suǒyǐ and yīncǐ (respectively, 60 and 40% of the 30 instances, Node 4). The third split of PropAtt (Node 4) set apart Physical act from Mental act/Fact. The 140 cases of Mental act/Fact (Node 5) contain equal amounts of yīncǐ and yúshì (both nearly 40%) and circa 25% of suǒyǐ. The bin with 166 instances of Physical act (Node 6) contains predominantly yúshì (60%) and roughly an equal amount of suǒyǐ and yīncǐ (both circa 20%). The predictive accuracy of this model was moderate (C = 0.53). In our case, there are four choices to be made (the four connectives). A random model would predict that each connective occurs in 25% of the observations. It can be concluded that 0.53 is a considerable improvement over 0.25.

The analysis shows that PropAtt played a role in three of the four significant splits in the data, suggesting that it was the most important predictor in the model. This interpretation was confirmed in Figure 7, which was the result of the random forest analysis. It gave information about the importance status of the features in distinguishing between the connectives under study. PropAtt stood out as the most important feature in the model. LingReal-SoC ranked second, followed by IdSoC.

Figure 7:

Evaluation of conditional importance of the variables PropAtt, IdSoC and LingReal-SoC in distinguishing between kějiàn, suǒyǐ, yīncǐ and yúshì.

5.1 Summary of the results

This study was carried out to answer three research questions. The first asks whether the connectives differ systematically with respect to their prototypical behaviors in term of subjectivity, the notion of which has been operationalized in terms of three features “the propositional attitude of the result segment”, “the identity of the SoC” and “the linguistic realization of the SoC”, and whether the prototypical profile of each connective remains robust across three different discourse types: naturally occurring conversations, microblog messages and newspaper texts. The second asks whether (and to what extent) the three discourse types differ with respect to these subjectivity features, as far as the causal construction is concerned. The third research question asks which of the three features is the most important in determining the profile of a particular connective.

To answer the first two questions, a systematic log-linear analysis was carried out for the relationship between connective, subjectivity and corpus. We have found two significant interactions connective*PropAtt and connective*IdSoC, which were not moderated by the factor corpus. The results strongly support the conclusion that the connectives have distinctive and robust subjectivity profiles regarding these two features.

There was only one set of findings that ran counter to this claim of robustness: the interaction connective*LingReal-SoC was found to be moderated by corpus (Section 4.3). This moderation was mostly related to the connectives kějiàn and yúshì (for details, see Table 8). Kějiàn instances had relatively rare explicit SoCs in newspaper and conversation data, while relatively many explicit SoCs in microblogs. This may well be a direct reflection of the property of the connective kějiàn, and the register of microblog discourse as well. It is part of the subjectivity profile of kějiàn to express opinions from the first-person perspective (as witnessed by the findings regarding PropAtt and IdSoC reported in the sections above). This profile fits well with the content of micro-blog posts, which is usually about the bloggers’ own experiences, personal feelings or opinions. They themselves are often the participants of the speech event. This would naturally lead to relatively many cases of explicit reference to the SoC. By contrast, in writing or conversational communication, it can be expected that the speaker SoC is often drawing conclusions or exchanging opinions about the speech event concerning a third-person participant. As a consequence, there might be less cases of explicit reference to the SoC in the utterance.

For yúshì, the occurrences of pro-drop were more frequent than expected in newspaper and (especially) microblog data. A possible explanation is the various degrees of “reviewability” (Clark and Brennan 1991: 141) of the discourse in each corpus. Both microblogs and newspapers consist of written texts. When the reference to the SoC is dropped out in the utterance, it is still retrievable from the context and therefore can be easily recovered by the addressee. In oral communication, however, messages are exchanged in a dynamic continuum. Frequent use of pro-drop might lead to an ambiguous delivery of the message or hamper the communication flow. Another corpus sensitive finding concerned the case of explicit SoC, which occurred more than expected only in newspaper discourse. This we think reveals the explicit character of written language (Finegan 1987), especially with regard to formal writing like newspaper articles.

Only with respect to the linguistic realization of the SoC, corpus turned out to be a factor affecting the behaviors of the connectives, and that relevance was limited. Overall, the subjectivity profiles of the connectives remain stable across corpora. To be precise, kějiàn displayed a highly subjective profile as it co-occurred predominantly with implicit SoC in each corpus. Yúshì, by contrast, is the most objective connective, featuring rarely any implicit SoC and mostly explicit or absent. Suǒyǐ and yīncǐ had a generic character with respect to this feature.

In sum, the study provides clear answers to RQ1. Kějiàn, suǒyǐ, yīncǐ, and yúshì as causal connectives are distinct from each other in terms of subjectivity and their specific subjectivity profile is robust across corpora. Kějiàn proves to be the most subjective connective, expressing predominantly judgments drawn by an implicit speaker SoC. Yúshì is the most objective connective, introducing prototypically factual consequents that are independent of an SoC, or intentional physical acts. The SoC that undertakes the physical act can be both the speaker and (significantly more often) the character who are mostly explicitly referred to (and are dropped in microblogs). Suǒyǐ and yīncǐ are relatively neutral, both co-occurring with all PropAtt categories and a majority of speaker SoCs, as well as a moderate amount of character SoCs and no SoC. Nevertheless, suǒyǐ appears to be slightly more subjective in that it expresses significantly more speech acts; yīncǐ, in comparison, has less cases of speaker SoC and more cases of no SoC. This leads to the following ordering of the connectives on a subjective-objective continuum, the order would be kějiàn>suǒyǐ >yīncǐ>yúshì.

The answer to RQ2 is that the variation between different discourse types with respect to subjectivity is also reflected at the level of the causal construction, as witnessed by the two significant interactions of corpus*PropAtt and corpus*IdSoC. The differences are relatively small. Microblog has relatively many speech acts; newspaper has relatively many character SoCs (which are rare in microblog). It seems that microblog discourse is the most subjective type, that newspaper discourse the most objective and that oral speech is in between.

The answer to RQ3 follows the findings from the conditional inference tree analysis and the random forest analysis (Section 4.4). The type of propositional attitudes expressed in the result segments proved to be the most important feature in characterizing the meaning and use of the connectives and making distinctions between them. It is involved in three of the four significant splits of our data. The linguistic realization of the SoC and the identity of the SoC follow in a decreasing order of importance.

5.2 Implications of the findings

The prototypical subjectivity profiles of the four RESULT connectives under analysis are to a large degree compatible with the conclusions drawn by previous studies (for details see Section 1). However, the robustness of the profiles of suǒyǐ and yīncǐ does not match the findings of a study using written data (Li et al. 2013). In that study, both connectives co-occur with more physical facts (counterpart of physical acts and facts in our model), explicit and character SoC in novels than in news reports and newspaper opinion pieces. As the present study did not include data from novels, there is no possibility for us to make direct comparisons. It could be argued that, compared to news reports and opinion pieces, the narratives in novels are more detached from the communicative here and now, and are more likely to be devoted to the description of the characters’ activities.

Another point worth mentioning is our finding that yīncǐ bears an objective flavor. It occurs infrequently with speaker SoC and tends to favor the most objective category no SoC (Table 7), which is indicative of the objective non-volitional relation. Nevertheless, it should be noted that the significance is relatively small based on our corpus data. It requires more empirical studies (for example, using experimental methods) for a clearer picture of this objective tendency.

By and large, our findings derived from the modified model are consistent across our dataset in each corpus. This consistency confirms the validity of the integrated subjectivity approach to the semantic-pragmatic properties that are prototypical of each connective. Moreover, through the systematic analysis of data from theoretically different discourse varieties, we have gained more insight into the use of Mandarin causal connectives, especially in the speech act domain, which has been substantially restricted in previous studies using written data. In the present study, a case in point is suǒyǐ, which is found to be preferred in the speech act domain (Table 5). This finding suggests that suǒyǐ, which prototypically has a relatively neutral profile, has a subjective flavor. In this sense it differs from yīncǐ, the other relatively neutral RESULT connective. The potential of a finer-grained distinction between these two connectives is promising.

Furthermore, the adapted model sets pro-drop apart as a separate category, as suggested in Xiao et al. (2021). This asks for further study concerning the subjectivity characteristics associated with this specific linguistic phenomenon. To the best of our knowledge, these two studies are the first to investigate the subjectivity characteristics of pro-drop. The model assumes that a dropped SoC is more subjective than an explicit SoC, but less subjective than an implicit SoC. Results show that pro-drop as a category rarely occurs in the relations with subjective kějiàn but mostly in its objective counterparts with yúshì. It may be that pro-drop fits best in a volitional content environment. This is obviously an issue for further study, not only on causal relations, but also other categories of coherence relations, and potentially not only restricted to Mandarin Chinese, but also considering other pro-drop languages.